Frames and the 3 dimensions of FNBr

Frames and Frame Semantics

Frame Semantics is the study of how linguistic forms evoke the cognitive structures — frames — that shape a speaker's understanding of them. Its basic assumption is that essentially all content words require, for their understanding, an appeal to a background frame within which the meaning they convey is motivated and interpreted. A word does not carry its meaning as an isolated atom; it points into an organized package of knowledge, and to understand the word is to recognize the relevance of that background.

The standard illustration is Langacker's hypotenuse: no definition of the term succeeds without invoking the right-angled triangle, since a hypotenuse just is the slanted side of that figure. The word profiles one element against a base that must already be in place. The same holds far beyond geometry. Tuesday presupposes the reckoning of time in a seven-day cycle; alimony presupposes divorce and the contracts entered into at its occasion; weekend presupposes a distinction between designated days of work and non-work. In each case the lexical unit — the pairing of a word with one of its meanings — evokes a frame and profiles some component of it, and the frame is the condition of the word's intelligibility.

Frames themselves are the general cognitive structures studied across the cognitive sciences under such names as schema, script, and idealized cognitive model: organized packages of knowledge, belief, and practice that let humans make sense of their experience. Some frames we owe to living on the earth, some to having human bodies, some to belonging to a particular culture, and some to being part of a particular speech community. They are typically representable as roles together with constraints on their fillers — the participants and props one can expect in a visit to a restaurant, the stages of a human life, the organization of a face. A lexical unit that evokes such a frame inherits its whole structure of roles and expectations.

FrameNet

FrameNet is the lexicographic project that turns this theory into a resource. Its aim is to document the range of semantic and syntactic combinatory possibilities — the valences — of each word in each of its senses, through corpus-based, computer-assisted annotation of example sentences. Where Frame Semantics is the account of how forms evoke frames, FrameNet is the disciplined recording of which frames a lexicon's words evoke and how those frames are realized in text.

Three features distinguish it. First, it proceeds frame by frame rather than word by word: progress is measured in frames completed, not entries alphabetized, and words are grouped by the scenes they share rather than by spelling. Second, it is built on frame elements — the roles a frame licenses (the Buyer, Seller, Goods, and Money of a commercial transaction; the Agent, Theme, and Goal of a placing) — which are annotated on attested corpus sentences so that each lexical unit's combinatorial profile is exhibited, not merely asserted. Third, it records a network of relations between frames — inheritance, perspective, subframe, using — so that the resource is not a flat list but a structured web in which a specific scene inherits from a general one and alternative construals of a single scene are held together.

This lexicographic core is what FrameNet Brasil extends the Berkeley framework to cover for Brazilian Portuguese, and it is the core that the dimensions described below are designed to serve rather than to replace.

FrameNet development (The book, ch. 2)

• All LUs in a Frame should have the same number and type of FEs, both explicitly and implicitly (NI).
• The basic semantic type of an FE should be constant in all uses of the LUs.
• When some of the target complement (LU) types are appropriate for the ontological category of the FE while others that are metonymically related are not, two distinct FEs are usually created, related to each other by the Excludes relation.
• In Frames that are aspectually complex, the LUs should imply the same set of stages and transitions.
• If the LUs differ with respect to whether typical and intended stages are actually achieved, the Frame should be split.
• The same FEs will be profiled by all LUs in a Frame, that is, the same participant viewpoint should be emphasized by all LUs.
• The interrelationships between the FEs of a Frame should be the same for all LUs.
• The assumptions, expectations, and concomitances of the LUs in a frame should be shared.
• The basic denotation of the LUs in a Frame should be similar.
• It is not possible to reliably base the distinction between Frames on the selection constraints of the LUs.
• The specifications that the LUs impose on the various FEs should be similar.
• The goal of all LUs in a Frame sharing the selection constraints in an FE is often relaxed in practice, as this would result in a very large number of frames.
• Words that are antonyms are grouped in the same frame.
• The ultimate test for these decisions is their usefulness.

Three dimensions

The current version of FrameNet Brasil is sketched as having three dimensions. They do not alter FrameNet's original goal as a lexicographic project — the annotation of lexical units against corpus evidence remains exactly what it was. What they add is a ground for frame understanding and frame creation: a set of reference structures that a modeller consults when placing a new frame in the resource, or when auditing an existing one, so that the analysis a frame encodes is made explicit and its consistency with the rest of the inventory becomes visible.

The dimensions are, throughout, descriptive and not prescriptive. As in FrameNet generally, and as in the DUL ontology the model borrows from, their job is to encode the modeller's analysis and surface inconsistencies, never to forbid a construal. Full coverage is not the target; useful, honest progress is. Each dimension is orthogonal to the others — it answers a different question about a frame — and each has its own reference document, of which this is the introduction. The three questions are:

1. the ontological dimension — what kind of thing is this, and which view does the frame take of it?
2. the conceptual-spaces dimension — what abstract structure does the frame's meaning rest on?
3. the domain dimension — what area of human experience is the frame about?

The ontological dimension

The ontological dimension classifies a frame and the lexical units it hosts by what they are and how they are construed. It rests on the DUL distinction between an Event — something that occurs, with one stable identity — and a Situation — a view of that event taken under some description. Erosion is one event; erosion-as-accomplishment and erosion-as-transition are two situations over the same event, neither of which changes the event's identity.

This maps onto two complementary layers. A coarse ontological type attaches to the lexical unit and tracks the identity of the concept — is it an event, an entity, a state, an attribute, or a relation? It exists to catch cross-class errors, such as an attribute concept lodged in an eventive frame. Above it, a set of namespaces attaches to the frame and records which view the frame takes over the underlying event — @phenomenon, @change, @stative, @agentive, and their kin. Because the type tracks identity and the namespace tracks the view, one type maps to many compatible namespaces, and a single event may legitimately be viewed under more than one of them: o gelo derreteu can be read as a change (the resulting liquid state) or as a phenomenon (the melting process). The model expects such pairs and never treats them as errors. This is FrameNet's native notion of perspective, applied at the level of the concept and its frame. The namespaces are described in the Namespaces document; the type inventory in Ontological Types for LU.

The conceptual-spaces dimension

The conceptual-spaces dimension supplies the abstract structure against which a frame's meaning is read. It is a layer of highly schematic frames — a FrameNet-native implementation of the image-schema tradition — grounded in Gärdenfors' theory of conceptual spaces. Where the image-schema tradition speaks of containment, sequence, part–whole, and force, this dimension realizes those recurrent structures as frames, using the existing frame apparatus rather than a parallel formalism. A lexical frame Uses a conceptual schema, and the schema supplies the structure the frame's meaning is read against; lexical frames that share a structure Use the same schema, so the schema is the generalization over them.

The layer is built from two primitives — REGION (any concept, modelled as a region in a conceptual space, whose location is its meaning) and FORCE (something that acts on regions, known only by its effect) — and from the single claim that every relation between regions is a force. Force has two effects: it can configure regions, holding them in place relative to one another (the static schema, DIMENSION, answering how things are), or it can transform a region, moving it from one position to another (the dynamic schema, EVENT, answering what happens). What appears static is force in equilibrium; what happens is force out of equilibrium. The layer introduces no new database mechanism — it lives entirely at the frame-element level, using coercion (FE-F) and the ternary rel_structure relation already in place — and, like the rest of the model, it keeps its inventory small and extends it empirically. It is detailed in the Conceptual Schemas document. (One caution recorded there: Gärdenfors' spaces are metric, whereas this apparatus is topological and relational — it records that regions are organized and how, not how far apart they are.)

The domain dimension

The domain dimension is the coverage taxonomy: it answers what area of human experience a frame is about. Seven fundamental domains — Physical, Biological, Social, Cultural, Psychological, Representational, and Moral — each reflect a distinct dimension of reality as human cognition and culture organize it, and each divides into subdomains, which may divide further. Every frame resolves to exactly one domain node, its most specific applicable home, either directly or by inheritance: a frame earns its own domain only when it introduces coverage its ancestors lack, and otherwise inherits its parent's. This gives full classification of the inventory — no frame left with a silent "no domain" status, structural and abstract frames included — while exposing which areas are richly covered and which are thin. The domain node answers only what area of human experience a frame is about; it does not model the internal structure of a scene, which is the business of the ontological dimension's namespaces. The seven domains and their subdomains are set out in the Domains document.

How the dimensions fit together

The three dimensions are independent axes over the same frame. Asking what a frame is (ontological), what abstract structure it rests on (conceptual spaces), and what it is about (domain) are separate questions with separate answers, and a frame carries a value on each. A change-of-state frame for melting, for instance, is a @change frame over an event-typed concept (ontological), Uses an EVENT schema whose endpoints range over a physical-state DIMENSION (conceptual spaces), and resolves to the Physical domain's Physical_transformation subdomain (domain). None of these classifications competes with the others, and none disturbs the lexicographic annotation of the frame's lexical units. Together they give a modeller a principled place to stand when understanding an existing frame or creating a new one — a reference structure to relate the frame to — while leaving the resource's original purpose, the corpus-grounded description of the lexicon, intact.